Support for GCC 15's incremental LTO

[thesamesam]

GCC 15 supports incremental LTO via -flto-incremental=/path/to/cache/dir (docs) for fast rebuilds with LTO with small source file changes.

Currently, we have:

• -Db_thinlto_cache
• -Db_thinlto_cache_dir

...which support roughly the same thing in Meson for Clang.

GCC defaults to "thin LTO" so tying it to -Db_lto_mode=thin doesn't make much sense unless that were to become a default, and would likely IMO propagate confusion.

But it seems a shame to provide duplicate options that do essentially the same thing. I think it was a mistake to use specific vendor technology in the naming & description of these options, but it is what it is.

What do we do going forward?

Suggested options (more welcome):

• Make -Db_thinlto_cache and -Db_thinlto_cache_dir do the right thing for GCC, and ignore the bad naming (this may be a bit messy given it only does something for Clang if -Db_lto_mode=thin)
• Add duplicate options representing the same thing for GCC
• Deprecate -Db_thinlto_cache and -Db_thinlto_cache_dir and replace them with new options that work for both Clang and GCC

[dcbaker]

Of those options I am strongly against option 2 (duplication), and could be convinced by either 1 or 3, but don't have a strong feeling on which is better.

On 3: we have this case, and we have the discussion around b_sanitizer with the list of sanitizers stuff, and I've always been a bit annoyed that we baked ndebug instead of something like asserts or debug_asserts which again falls into the "we baked specific terminology into Meson". We have machinery for handling deprecated options, with the exception of get_option which just becomes a bunch of open coding in the interpreter. So maybe it's time to really address the issue of "what to do about changed/renamed/deprecated builtin options"?

[dcbaker]

Related, I had looked at having switches for rust LTO, which has thin and fat support, but also a per-target LTO that doesn't cross target boundaries (so all of the .rs files in liba.a get LTO, and all of the .rs files in libb.a get LTO, but liba and libb don't get LTO when they're linked together)

https://doc.rust-lang.org/rustc/codegen-options/index.html#lto

[thesamesam]

Looks like #12617 actually adds some of that machinery for us wrt deprecation.

[thesamesam]

cc @bonzini